PreprintPDF Available

Augmented reality for supporting the interaction between pedestrians and automated vehicles: An experimental outdoor study

Authors:

Abstract

Communication from automated vehicles (AVs) to pedestrians using augmented reality (AR) could positively contribute to traffic safety. However, previous AR research for pedestrians was mainly conducted through online questionnaires or experiments in virtual environments instead of real ones. In this study, 28 participants conducted trials outdoors with an approaching AV and were supported by four different AR interfaces. The AR experience was created by having participants wear a Varjo XR-3 headset with see-through functionality, with the AV and AR elements virtually overlaid onto the real environment. The AR interfaces were vehicle-locked (Planes on vehicle), world-locked (Fixed pedestrian lights, Virtual fence), or head-locked (Pedestrian lights HUD). Participants had to hold down a button when they felt it was safe to cross, and their opinions were obtained through rating scales, interviews, and a questionnaire. The results showed that participants had a subjective preference for AR interfaces over no AR interface. Furthermore, the Pedestrian Lights HUD was more effective than no AR interface in a statistically significant manner, as it led to participants more frequently keeping the button pressed. The Fixed pedestrian lights scored lower than the other interfaces, presumably due to low saliency and the fact that participants had to visually identify both this AR interface and the AV. In conclusion, while users favour AR in AV-pedestrian interactions over no AR, its effectiveness depends on design factors like location, visibility, and visual attention demands. In conclusion, this work provides important insights into the use of AR outdoors. The findings illustrate that, in these circumstances, a clear and easily interpretable AR interface is of key importance.
Augmented reality for supporting the interaction
between pedestrians and automated vehicles: An
experimental outdoor study
31 December 2023
Thomas K. Aleva*, Wilbert Tabone*, Dimitra Dodou, Joost C.F. de Winter
Faculty of Mechanical, Maritime and Materials Engineering, Delft University of Technology
* Joint first authors
Corresponding author: j.c.f.dewinter@tudelft.nl
Abstract
Communication from automated vehicles (AVs) to pedestrians using augmented reality (AR)
could positively contribute to traffic safety. However, previous AR research for pedestrians was
mainly conducted through online questionnaires or experiments in virtual environments instead
of real ones. In this study, 28 participants conducted trials outdoors with an approaching AV and
were supported by four different AR interfaces. The AR experience was created by having
participants wear a Varjo XR-3 headset with see-through functionality, with the AV and AR
elements virtually overlaid onto the real environment. The AR interfaces were vehicle-locked
(Planes on vehicle), world-locked (Fixed pedestrian lights, Virtual fence), or head-locked
(Pedestrian lights HUD). Participants had to hold down a button when they felt it was safe to
cross, and their opinions were obtained through rating scales, interviews, and a questionnaire.
The results showed that participants had a subjective preference for AR interfaces over no AR
interface. Furthermore, the Pedestrian Lights HUD was more effective than no AR interface in a
statistically significant manner, as it led to participants more frequently keeping the button
pressed. The Fixed pedestrian lights scored lower than the other interfaces, presumably due to
low saliency and the fact that participants had to visually identify both this AR interface and the
AV. In conclusion, while users favour AR in AV-pedestrian interactions over no AR, its
effectiveness depends on design factors like location, visibility, and visual attention demands.
In conclusion, this work provides important insights into the use of AR outdoors. The findings
illustrate that, in these circumstances, a clear and easily interpretable AR interface is of key
importance.
Keywords: Augmented reality; Pedestrian safety; Anchoring; See-through AR
Introduction
Every year, 300,000 pedestrian deaths occur worldwide, accounting for 23% of all road fatalities
(World Health Organization, 2018). The United Nations has set a target within its Sustainable
Development Goals to halve the number of traffic deaths by 2030 (United Nations, 2020). In this
resolution (A/RES/74/299), it was also noted that “continuous progress of automotive and digital
technologies could improve road safety, including through the progressive development of highly
and fully automated vehicles in road traffic” (United Nations, 2020, p. 4).
One possibility to improve the safety of pedestrians is to equip automated vehicles (AVs) with
better sensors and intelligence so that they can respond earlier, in order to reduce collisions with
vulnerable road users (Jungmann et al., 2020; Paiva et al., 2021). Another possibility is the
introduction of AV-to-pedestrian communication technologies. Previous research has shown that
with displays on the outside of the AV, called external human-machine interfaces or eHMIs,
pedestrians can make more effective road-crossing decisions (Bazilinskyy et al., 2021;
Bindschädel et al., 2022; Dey et al., 2020). However, eHMIs have certain disadvantages in that
they typically cannot address an individual pedestrian (Colley et al., 2020; Tran et al., 2023) and
1
that they can be difficult to perceive in some cases, for example due to an occlusion by another
object (Dey et al., 2022; Troel-Madec et al., 2019).
Recently, AV-to-pedestrian communication via wearable devices has been proposed (Gelbal et
al., 2023; Hasan & Hasan, 2022; Lakhdhir et al., 2023). More specifically, augmented reality (AR),
defined as a technology that overlays virtual information onto the real world, appears to be a
promising solution for communicating with pedestrians (Tabone et al., 2021b). On the one hand,
the use of AR might sound unusual and undesirable; after all, it may be questioned whether
pedestrians would want to rely on an expensive headset in order to move safely through traffic
(e.g., Berge et al., 2022; Tabone et al., 2021a). At the same time, given that other AR technologies
such as car head-up displays (HUDs) (Transparency Market Research, 2023) and Google Maps
visual overlays on mobile phones (Google, 2020), once seen as futuristic, are now common, it is
conceivable that in the future, pedestrians will receive visual assistance through AR glasses.
Although AR is a potentially promising technology for pedestrians, still little research exists on this
topic. One of the problems is that AR is currently challenging to implement, with only the most
recent headsets capable of creating a compelling and user-friendly experience (e.g., Microsoft
Hololens 2, Varjo XR 3, Magic Leap 2). While exceptions exist (e.g., Kang et al., 2023; Zhang et
al., 2023), much previous AR research for pedestrians has limitations: some studies focus solely
on the orienting phase without any form of human-subject evaluation (Tabone et al., 2021b; Tong
et al., 2021), others use questionnaires with photos or videos (Hesenius et al., 2018; Tabone et
al., 2023a; Wilbrink et al., 2023), and still others test AR concepts in immersive virtual reality (VR)
environments (Malik et al., 2023; Pratticò et al., 2021; Tabone et al., 2023b; Tran et al., 2022). A
critical note that should be made in this last point is that—although valid results can be obtained
in VR—this is strictly speaking not AR. After all, merely adding virtual displays within a completely
virtual environment is still considered VR, rather than AR. AR is different from VR in that AR
incorporates elements of the physical world into the user’s experience, creating a blend of virtual
and real-world elements (Rauschnabel et al., 2022).
In this study, we share the findings from an outdoor experiment involving participants using an
AR headset. In addition to the AR interfaces, the approaching AV was also simulated. The use of
a virtual AV, as opposed to a real one, ensured consistent experimental conditions and simplified
the process of attaching AR elements to the moving AV. Our approach serves as a stepping stone
to a future full AR experience, in which pedestrians can walk around untethered while the AR
headset wirelessly communicates with an actual AV. Our use of AR, where virtual road users
coexist with real humans in an outdoor space, has been applied by several others before, though
not in the context of AV-pedestrian communication (Kamalasanan et al., 2022; Maruhn et al.,
2020). Our methodology also corresponds with other realistic simulation methods, such as
displaying virtual road users in a real car (Bokc et al., 2007; Butenuth et al., 2017; Feng et al.,
2018; Hussain et al., 2013; Sheridan, 2016), a technique that supersedes driving simulators which
inherently offer limited visual and motion cues. In our approach, while the environment remains
the real world, it is enhanced with simulated experimental stimuli.
In the current experiment, four different AR interfaces and a no-AR-interface baseline condition
were compared among human participants. The focus was on whether these AR interfaces made
participants feel safe to cross, as well as subjective qualities including whether the AR interface
was found to be intuitive and was accepted. The interfaces used in this study were adopted from
three prior works: an AR concept design study (Tabone et al., 2021b), an online questionnaire
with video clips in which participants provided subjective ratings (Tabone et al., 2023a), and a
CAVE-based simulator study in which participants crossed a virtual road (Tabone et al., 2023b).
Our goal was to investigate whether the results we found in the current experiment correspond
with the earlier research among human participants (Tabone et al., 2023a, 2023b).
The tested AR interfaces differed in terms of their anchoring techniques: the AR interface was
either vehicle-locked, world-locked, or head-locked, three methods that bring fundamentally
different demands regarding the user’s visual attention (Lebeck et al., 2017; Lingam et al., 2023;
Peereboom et al., 2023; Tabone et al., 2023b). A vehicle-locked AR interface entails, just like an
eHMI, that implicit communication in the form of the speed of and distance to the AV and explicit
AR cues are congruent in time and place. A head-locked AR interface, on the other hand, is
always visible but requires that the AV is looked at to confirm the cues of the AR interface. Finally,
2
a world-locked interface requires attention to be distributed between the AV and the AR interface,
with the possible advantage that the AR interface can be presented at a fixed and familiar location.
For illustration, Lingam et al. (2023) conducted a study using a VR driving simulator to investigate
whether eHMIs were better positioned on the roof of an approaching AV, or integrated into the
road infrastructure, resembling a traffic light configuration. Their results showed that both
solutions were more favoured by the car drivers than the absence of any signalling. However,
each concept presented its own set of advantages and disadvantages. The advantage of the
infrastructure solution (i.e., world-locked signal) was that the signals were positioned at a fixed
and accessible location before the intersection, and the driver could see this from a distance. An
eHMI on the AV (i.e., vehicle-locked signal), on the other hand, required that the driver first turn
their head/eyes to see the AV with its eHMI; once identified, however, the driver could directly
infer what the AV was going to do, as deduced from its speed and the eHMI signal.
The current study, conducted with pedestrians outdoors instead of in VR, aimed to determine
whether similar strengths and weaknesses could be identified for various AR anchoring methods.
We implement four AR interface concepts, taken from a CAVE study by Tabone et al. (2023b):
(1) Virtual fence, a world-locked interface that is highly visible, yet may impart a false sense of
security and which partially occludes the environment due to its semi-transparent walls, (2) Fixed
pedestrian lights, another world-locked interface; it has a familiar design and is smaller in stature
compared to the Virtual Fence, (3) Pedestrian lights HUD, a head-locked interface similar to a
conventional traffic light, but always visible in the users field of view, and (4) Planes on vehicle, a
vehicle-locked AR interface similar to an eHMI. We adopted the designs from Tabone et al.
(2023a, 2023b) without design improvements. Note that several insurmountable differences exist
between the AR implementation of this study and that of the online and VR studies Tabone et al.,
such as in timing and visibility in the external environment, thereby making an exact numerical
comparison not our objective. Instead, we opted for a qualitative comparison, focusing on the
effects of the anchoring method. In addition to comparing different AR interfaces with no AR
interface, one of the objectives of this research was to document, accompanied by Unity source
code, how AR research in the outdoor environment can be conducted.
Methods
Participants
Participants were recruited through personal networks, without the offer of financial
reimbursement or other incentives. The study included a total of 28 persons, comprising 23 males
and 5 females, with ages ranging from 19 to 59 years. The average age was 27.2 years, with a
standard deviation of 9.2 years. In response to the intake questionnaire item ‘Do you use any
seeing aids?’, 19 participants answered ‘no’, 5 answered ’yes, glasses’, and 4 answered ‘yes,
contact lenses’. We did not record whether participants wore glasses or contact lenses during the
experiment. However, it was observed that at least one participant wore their glasses under the
headset, without noticeable problems.
The intake questionnaire further indicated that 19 of the participants were students. Twenty-six
participants indicated being Dutch, one Chinese, and the nationality of one participant was
unknown. Twenty-six participants held a driver’s licence, for an average of 8.0 (SD = 7.3) years.
A total of 17 participants had used a VR headset before while 9 participants had used an AR
headset before. Daily walking time was less than 15 min for 7 participants, 15 to 30 minutes for
12 participants, and more than 30 min for 9 participants. Cycling was the primary mode of
transportation for most participants (22 out of 28). A test for colour blindness showed that one
participant was colorblind. This person was not excluded from participation, since the AR
interfaces featured redundancy gain in that they did not solely rely on red/green colours, but also
used icons, movement, or stimulus location cues. Additionally, the inclusion of this participant was
considered beneficial for gaining valuable insights.
Each participant provided written informed consent. The experiment procedure was approved by
the TU Delft Human Research Ethics Committee, approval no. 3054.
Materials and Settings
The experiment took place in a designated area of the Delft University of Technology campus,
which was closed to normal traffic. It was conducted on an Alienware PC powered by an Intel®
3
Core™i7-9700K CPU at 3.60 GHz, equipped with 16 GB RAM and a Nvidia GeForce RTX 2080
Ti GPU. The AR software ran in Unity 2021.3.13f1 (Unity, 2022), combined with the Varjo XR
Plugin (Varjo Developer, 2023). A custom script was used that allowed the experimenter to select
the experiment conditions from within Unity.
AR was displayed by means of a Varjo XR-3 headset. The Varjo XR-3 provides a 90 Hz refresh
rate and a 115° horizontal field of view. The focus area, of 27° × 27°, was rendered at 70 pixels
per degree on a µOLED display, providing 1920 × 1920 pixels per eye. The peripheral area was
rendered at about 30 pixels per degree on an LCD, producing 2880 × 2720 pixels per eye. A pole
was used to route the cables from the PC to the participant (Figure 1).
Figure 1. Participant wearing the tethered Varjo XR-3 headset.
The Varjo XR-3 presented a virtual AV and virtual AR interfaces while depicting the real world by
means of video pass-through. Within the headset configuration software ‘Varjo Base’, the imaging
and exposure settings were set to automatic, and the highest image quality settings were
selected. This proved to be feasible for our application without glitches or delays in data storage
or visual animations.
Our initial intention was to anchor the virtual objects to the real world by means of object tracking
and reference markers (Varjo Technologies, 2023). However, this approach proved to be
unreliable in our outdoor setting, presumably due to the relatively empty environment and sunlight
reflections on the markers. Instead, tracking for the headset was adopted, by means of two
SteamVR Base Stations 2.0 (HTC VIVE, 2019), as depicted in Figure 1. Our solution was to
position the AR Origin at a fixed position, based on the initial calibration position and orientation
of the Varjo XR-3. Each time the hardware was set up, calibration of the HMD tracking was
performed first. In Unity, an invisible Ground place was created, which functioned as a surface
for the AV to drive on. The scale value of the AR Origin was set such that distances in the virtual
world corresponded with distances in the real world. Unity was configured to allow participants to
rotate their heads and look around, but not to translate or move through the environment. The
4
height of the Main Camera above the ground plane was fixed at 1.70 m. Our outside-in tracking
approach provides a simple and robust way to present virtual overlays onto the real world.
However, the location of these overlays is not anchored to the real world, and can somewhat vary
depending on the precise calibration of the setup.
A Logitech R400 wireless presenter was used as the remote control. The right arrow button was
used by the participant as the button to indicate if the participant felt it was safe to cross. The
receiver was connected via a USB port of the Alienware desktop.
The sound of the AV was transmitted through headphones that were plugged into the Varjo XR-
3. The headphones did not include noise-cancelling, so surrounding real-world environment noise
could still be heard.
Participant Information and Task
Participants were provided with information about the procedure and tasks through a leaflet. They
were informed that they would wear an AR headset, which uses cameras to capture the
environment and can project virtual objects onto this real-world setting.
Participants were informed that they would stand at the side of a road and that their task was to
press and hold the button when they felt it was safe to cross and to release it when they did not
feel safe to cross. It was emphasised that participants should not physically cross the road. This
method for measuring crossing intentions was previously introduced by De Clercq et al. (2019)
and was considered favourable as it allows participants to respond almost immediately to
changing conditions. In other research, we had participants actually cross (Tabone et al., 2023b)
or communicate their crossing intention or ‘critical gap’ through hand gestures (Rodríguez
Palmeiro et al., 2018) or by taking a single step (Epke et al., 2021). A drawback of these methods
is the variability among participants in executing these actions, making it relatively challenging to
extract their intentions or perceptions.
The leaflet also mentioned that, at the start of each trial, participants would see a circle (hereafter
named ‘attention-attractor circle’) at one of three locations, i.e., either in front of them, to their left,
or to their right. Participants were instructed to gaze at the circle for a duration of one second to
initiate the trial. It was mentioned that a virtual vehicle would approach from the right, potentially
stopping for the participant and possibly communicating this intent through various
communication interfaces. Finally, it was mentioned that, after each trial, participants would be
prompted with a question displayed in the AR environment, to which they were to provide a verbal
response, and that this procedure would be repeated for a total of four different AR interfaces,
plus a no-interface baseline condition. Participants were also informed that, following every block
with a particular AR condition, the experimenter would ask open-ended questions about their
experiences with the presented condition.
In previous research, the attention-attractor circles were used to impose visual distraction. This
was done to examine whether different AR interfaces exhibit different robustness to distraction
while the AR interface was active (Tabone et al., 2023b). Amongst others, Tabone et al. (2023b)
found that the Pedestrian lights HUD supported crossing decisions even though the participants
were visually distracted and had not glanced at the AV yet, because the HUD moved with the
participant’s field of view . In the current study, the circles played a different role. In our case, the
circles disappeared before the start of the trial, while the AR interface appeared 3.1 s later, which
allowed participants to direct attention to the approaching AV before the activation of the AR
interface. The purpose of the circles, in our case, was to reduce monotony of the task and to
increase visual demands.
Automated Vehicle Behaviour
In each trial, a virtual AV approached from the right, displaying one of two behaviours: non-
yielding or yielding. The vehicle model asset had dimensions of 4.95 m in length, 2.10 m in width,
and 1.35 m in height (Final Form Studio, 2021). The AV reflected light from a virtually positioned
sun, and the wheels rotated in accordance with the forward speed. Shadows were not simulated.
Audio was incorporated to produce a speed-sensitive engine sound, with a Doppler effect applied.
5
When the trial began, the AV appeared approximately 45 m away from the participant, with the
distance measured parallel to the road. Simultaneously, an attention-attractor stimulus in the form
of a cyan circle ring appeared either in front of, to the left of, or to the right of the participant. After
the participant looked at the circle for 1 s, the AV began driving at a speed of 30 km/h.
Once 3.1 s had elapsed from the moment that the AV began driving, it passed an invisible trigger,
prompting the AR interface to appear. In the yielding condition, the AV began decelerating 3.9 s
from the start of its movement and came to a complete stop at an elapsed time of 7.8 s. The AV
featured a slight forward pitch angle as it decelerated. In the non-yielding condition, the AV
maintained a speed of 30 km/h and passed the participant 6.5 s from the beginning of its
movement. Figure 2 provides an aerial view of the roadway, illustrating the sequence of events:
the AV starting, the appearance of the AR interface, the point where the AV began decelerating,
and where it came to a full stop.
Figure 2. Top-down view of the experiment area (overlay drawn on an image from Google Earth, 2023).
Augmented Reality Designs
The AR interfaces were adopted from earlier research in which these interfaces had been
designed (Tabone et al., 2021b), and after minor modifications, presented to a large sample of
online respondents (Tabone et al., 2023a), and subjected to experimental testing in an immersive
virtual simulator (Tabone et al., 2023b). The nine AR interfaces were previously divided into three
anchoring methods: world-locked, head-locked, and vehicle-locked (Tabone et al., 2023b). The
present study focused on comparing four selected AR interfaces. Specifically, from each
anchoring category, one was selected: Virtual fence (world-locked), Pedestrian lights HUD (head-
locked), and Planes on vehicle (vehicle-locked). In addition, the Fixed pedestrian lights interface
was selected. The reason for including this world-locked interface was to allow a comparison with
the Pedestrian lights HUD. The four selected AR interfaces were as follows:
1. The Virtual fence consisted of a zebra crossing projected on the road combined with
semi-translucent walls surrounding the zebra (Figure 3). The interface was 3.0 m tall, 2.5
m wide and 7.5 m long. A semi-translucent gate positioned in front of the pedestrian
opened in 1-s time after the interface appeared in the yielding condition, and remained
closed in the non-yielding condition.
2. The Fixed pedestrian lights resembled existing traffic lights (Figure 4). It remained
stationary across the street from where the pedestrian was positioned. The interface was
2.1 m tall and positioned 14 m from the pedestrian. It consisted of a pole with a box (0.20
m wide, 0.38 m tall) on top displaying either a lit-up red icon of a standing pedestrian, or
a green icon of a walking pedestrian.
3. The Pedestrian lights HUD was similar to the Fixed pedestrian lights interface (Figure 5).
However, instead of being world-locked, the same box was anchored to the user’s head.
It was placed at 1 m distance from the participant’s head and off-centre by 30 cm upwards
and to the right. The interface was rotated around the vertical axis in order to face the
user.
4. The Planes on vehicle was a vehicle-mapped interface consisting of a red plane with a
stop-hand icon or a green plane with an icon depicting a pedestrian crossing a zebra
6
crosswalk (Figure 6). The 2.1 wide and 1.58 m tall plane hovered above the front of the
AV so that the bottom of the plane aligned with the front bumper; it was tilted by 54° to
be positioned parallel with the AV’s windshield.
In all four AR interfaces, colour was used to provide a redundant cue; for the non-yielding AVs
this was pure red, and for yielding AVs this was pure green. The Virtual fence and Planes on
vehicle were semi-transparent, to ensure that the AV or possibly other relevant objects remained
visible to the participants.
Figure 3. Virtual fence in the non-yielding and yielding condition.
Figure 4. Fixed pedestrian lights in the non-yielding and yielding condition.
7
Figure 5. Pedestrian lights HUD in the non-yielding and yielding condition.
Figure 6. Planes on vehicle in the non-yielding and yielding condition.
Experiment Design
The experiment was of a within-subject design. Each participant was exposed to the four AR
interfaces and a no-interface baseline condition, two AV behaviours (yielding and non-yielding),
and three attention-attractor locations (left, middle, right).
Participants first completed two practice trials in the Baseline condition, one trial with a yielding
AV one trial with a non-yielding AV. Each of the five interface conditions was presented in a
separate block. Each block consisted of six trials: three with a yielding AV and three with a non-
yielding AV. The three trials were conducted with the attention attractor at the left, right, or middle.
This resulted in a total of 30 trials per participant (5 interface conditions × 2 yielding behaviours ×
3 attention-attractor locations). The order of the five blocks, as well as the order of the six trials
within each block, were counterbalanced using a Latin Square method.
The experiment lasted 45 to 60 minutes per participant, with the time spent wearing the headset
amounting to approximately 30 minutes.
Questionnaires and Rating Scales
After signing the consent form, participants completed an intake questionnaire, designed in
Qualtrics (Qualtrics XM, 2023), to gather demographic information (see Data Repository
8
Material). The intake questionnaire also included items about affinity for technology (ATI scale,
Franke et al., 2019), items about whether participants had experienced VR or AR headsets
before, and a brief test for colour blindness (Ishihara, 1917). Some participants had completed
the pre-experiment questionnaire before their scheduled experiment slot.
Next, participants were asked to read a leaflet describing a short description of the experiment. A
complementary oral explanation of the experiment was provided where needed. Subsequently,
participants put on the Varjo XR-3 headset, were handed the remote button, and a multi-point
eye-tracker calibration was conducted. After each trial, at exactly 10 seconds after the AV had
started moving, a statement appeared in front of the participant: This interface/situation was
intuitive for signalling: ‘Please do cross the road’.” for the yielding condition, or This
interface/situation was intuitive for signalling: ‘Please do not cross the road’.” for the non-yielding
condition. Participants verbally indicated to what extent they agreed on a scale from Fully disagree
(1) to Fully agree (7).
After each block of six trials with a particular AR condition, a semi-structured interview was
conducted regarding interface design qualities, the timing of the interface appearance, the
preference between yielding and non-yielding state, and the participant’s wellbeing according to
the misery scale (MISC; Bos, 2015).
After all five blocks, participants completed a post-experiment questionnaire in Qualtrics. This
questionnaire contained items related to the AR experience and about AR interfaces in general.
Participants were also asked to rank the five interface conditions in terms of their preference.
Additionally, they were asked, for the four AR interfaces, to answer items regarding the
intuitiveness of the green and red interfaces, convincingness of the green and red interface,
interface trustworthiness, size (too small, too large), timing (too early, too late),
clarity/understandability, and visual attractiveness, as well as a 9-item acceptance scale (Van der
Laan et al., 1997), identical to Tabone et al. (2023a). The items used 7-point scales, except for
the acceptance scale and adoption questions which use a 5-point semantic differential scale, and
the ranking item. Open questions were also asked per interface condition to allow participants to
justify their responses.
Data Recording and Analysis
The data was stored at a frequency of 50 Hz. Firstly, we determined per trial what percentage of
the time participants kept the response button pressed. This was done from the moment the AR
interface appeared until the vehicle came to a stop (yielding AVs) or passed (non-yielding AVs).
Additionally, from the post-experiment questionnaire, we determined a composite score per AR
interface, identical to how it was done by Tabone et al. (2023a). We calculated a composite score
because the self-reports were strongly correlated, and we found no evidence of multiple
underlying constructs (Tabone et al., 2023a). This composite score was calculated based on
participants’ responses to 15 items, which included a 9-item acceptance scale and 6 additional
items (1. intuitiveness for non-yielding AVs, 2. intuitiveness for yielding AVs, 3. convincingness
for non-yielding AVs, 4. convincingness for yielding AVs, 5. clarity/understandability, 6.
attractiveness). To calculate the composite score, we first concatenated the questionnaire results
from the 28 participants across 4 AR conditions, resulting in a grand 112 × 15 matrix. We then
standardised each of the 15 variables, so their mean became 0 and their standard deviation
became 1. Next, we summed the standardised scores of the 15 items, resulting in a 112-element
vector of total scores. These 112 scores were standardised again, providing a composite score
for each participant and each AR interface. The preference rank, in which participants had to sort
the five AR conditions from 1 (most preferred) to 5 (least preferred), was analysed as a separate
item.
Finally, the post-trial intuitiveness ratings were averaged over 3 trials per participant so that for
each AR interface and participant, and for both yielding and non-yielding AV, an intuitiveness
score was available. Since the location of the attention-attractor circles did not appear to have a
major influence on how the four AR interfaces were responded to, the results for the three trials
per AR interface condition and yielding condition were averaged. In the statistical analyses,
repeated-measures ANOVAs were used, with the AR condition as the independent variable. The
findings are shown as means, complemented by 95% confidence intervals for within-subjects
9
designs (Morey, 2008). For each of the nine measures, post-hoc comparisons between conditions
were conducted using paired-samples t-tests. An alpha value of 0.005 was chosen, which is more
conservative than the usual 0.05, due to a maximum of 10 possible combinations of AR conditions
that can be compared (4 + 3 + 2 + 1).
The interview results for each AR interface and participant were condensed into brief highlights
by one of the authors (the experimenter). These highlights from all 28 participants were
subsequently analysed and summarised per AR condition by another author using thematic
analysis, with a focus on the clarity of the AR interface and participant satisfaction. The resulting
summaries were checked with the interview transcripts (available for 16 out of the 28 participants)
to ensure representativeness and were adjusted as needed. For each AR interface, one or two
quotations were selected for presentation alongside the summary.
Results
The 28 participants performed a total of 840 trials. 4 of 840 log files were unavailable or could not
be used due to an experimenter error. For the post-experiment questionnaire, responses for the
Planes on vehicle were unavailable for 1 of 28 participants.
Misery scores were low, with 26 of 28 participants reporting no symptoms (0 or 1 on the 11-point
scale) and 2 of 28 participants reporting maximum scores of 3 and 4, respectively, at any time
during the experiment. One of these two participants indicated feeling ill already before the
experiment.
Figure 7 shows the percentage of trials in which participants held down the response button.
Interestingly, for both non-yielding AVs (Figure 7, left) and yielding AVs (Figure 7, right), a large
portion of the participants (approximately 60%) did not hold down the response button at the
beginning of the trial. The reasons for this are not clear, but it may be related to participants
experiencing high workload or the participants not being reminded to keep the button pressed at
the start of the trial. A technological explanation, involving the wireless connection of the button
device, cannot be ruled out either. Therefore, the results in Figure 7 should be considered in a
relative context, comparing the five conditions with each other, rather than in absolute terms.
Figure 7. Button press percentage for the four AR interfaces and baseline condition, for non-yielding AVs
(left), and yielding AVs (right). The grey background represents the interval between the moment the AR
interface appeared (at 3.1 s) and the AV passed (non-yielding AVs, 6.5 s) or the AV came to a full stop
(yielding AVs, 7.8 s).
For the non-yielding AV (Figure 7, left), participants, on average, released the response button
as the AV came closer, indicating that they felt less and less safe to cross. However, about 13%
of the participants did not have the response button released while the vehicle passed.
10
For the yielding AVs (Figure 7, right), participants started pressing the response button after the
AR interface appeared, at 3.1 s. For the baseline condition, this happened somewhat later, which
can be explained by the fact that the AV only started to slow down at 3.9 s (thus, 0.8 s after the
AR interfaces became visible). That is, in the baseline condition, participants could not anticipate
that the AV was going to stop until it actually started to decelerate. From the button press
percentages in the grey interval, it can be seen that the four AR interfaces were slightly more
effective than the baseline condition.
Figure 8 shows the button press percentages in the selected time interval for non-yielding AVs
(a) and yielding AVs (b), as well as the mean composite scores calculated from 15 items of the
post-experiment questionnaire (c), intuitiveness scores measured after each trial for non-yielding
AVs (d) and yielding AVs (e), the preference rank (f), trust (g), timing (h), and size (i). The green
numbers indicate pairs of conditions that exhibit statistically significant differences following a
paired-samples t-test with Bonferroni correction. Full results from the paired-samples t-tests are
accessible in the Data Repository.
The results presented in Figure 8 indicate that the baseline condition scored relatively poorly, as
demonstrated by a low button press percentage for yielding AVs (b), low intuitiveness scores
especially for non-yielding AVs (d) and yielding AVs (e), and on average the worst preference
rank (f).
Among the four AR interfaces, there were no major differences, although the Fixed pedestrian
lights scored relatively poorly. This is evident from the low composite score (c), low intuitiveness
scores (d & e), and the poor preference rank (f) among the four AR interfaces. However, none of
these effects were statistically significant. The largest effects between the four interface
conditions were, however, observed regarding the perception of the physical variables timing and
especially size. More specifically the Fixed pedestrian lights were perceived as clearly too small
(Figure 8i), and related to this, participants also believed that the traffic lights became visible too
late (Figure 8h); its timing was identical to that of the other three interfaces, but this assessment
might be due to the fact that it took participants extra time to visually locate it, creating the
impression that it became visible too late.
Figure 8. Means and 95% confidence intervals of participants’ (a) button press percentage for non-yielding
AVs (%), (b) button press percentage for yielding AVs (%), (c) composite scores based on post-experiment
self-reports (z-score), (d) post-trial intuitiveness scores for non-yielding AVs (1: fully disagree, 7: fully agree),
(e) post-trial intuitiveness scores for yielding AVs (1: fully disagree, 7: fully agree), (f) preference rank (1:
most preferred, 5: least preferred), (g) trust for decision-making (1: fully disagree, 7: fully agree), (h) interface
trigger timing (1: too early, 7: too late), and (i) interface size (1: too small, 7: too large). The numbers in green
11
indicate the conditions with which this condition shows a statistically significant difference (p < 0.005),
according to paired-samples t-tests.
The results of the repeated-measures ANOVAs corresponding to the data presented in Figure 8
are shown in Table 1. The effect sizes (partial η2) were not particularly strong for the button
presses (a & b), the composite score (c), or the trust score (g). However, they were fairly strong
and statistically significant for the post-trial intuitiveness ratings (d & e), the preference rank (f),
and as mentioned above, the timing (h) and size (i).
Table 1
Results of repeated-measures ANOVAs for the AR interface comparisons shown in Figure 8.
Dependent measure Measurement
moment
Df F p partial η
2
(a) Button press percentage, non-yielding AVs During trials 4,108 0.57 0.684 0.02
(b)
Button press
percentage
, yielding AVs
4,108
2.33
0.060
0.08
(c)
Composite score
Post
-
experiment
3,78
1.65
0.185
0.06
(d)
Intuitiveness, non
-
yielding AVs
Post
-
trial
4,108
2.88
0.026
0.10
(e) Intuitiveness, yielding AVs Post trial 4,108 6.34 < 0.001 0.19
(f) Preference rank Post-experiment
4,108 5.72 < 0.001 0.17
(g) Trust Post-experiment
3,78 0.91 0.441 0.03
(h)
Timing
Post
-
experiment
3,78
6.90
< 0.001
0.21
(i)
Size
Post
-
experiment
3,78
19.1
< 0.001
0.42
Note. For (c), (g), (h), and (i), data for Planes on vehicle were missing for one participant.
The results of the post-block interviews are indicated below:
Baseline. Without an interface, many participants felt the situation was more akin to real-
life scenarios. They indicated that the absence of clear signals, or feelings of insecurity,
without the interface caused them to take longer in making decisions. The factors
influencing their decision-making were the AVs deceleration and speed. Additionally, the
changing pitch of the AV’s sound during deceleration was reported to be a relevant cue.
“Well, at some point you see that car slowing down. Then you still remain a bit
apprehensive, thinking, ‘Okay, it's really stopping.’ And only then do I decide to start
walking.”
Virtual fence. Participants mentioned that the size of the Virtual Fence made it stand out,
and some even found it somewhat intimidating. The interface was generally perceived as
clear, particularly the green signal. However, the red signal was confusing for some
participants, who were unsure whether it warned them not to cross or indicated that the
AV would stop. The zebra crossing within the interface also posed a dilemma: it seemed
to encourage walking, but this was in conflict with the red walls. Additionally, some
participants expressed concerns that the interface might obstruct the view of other
potential road users.
Yes, so I found it a bit difficult at first, because you initially have those walls, which are
red. So then you think, oh, is it red for them? And are you protected by those walls or
something? And you're also in a zebra crossing. When I see a zebra crossing, I think, oh,
I need to walk. So I found that a bit counterintuitive. But once you get used to it, I think
you see it faster.
Fixed pedestrian lights. The Fixed pedestrian lights were a familiar concept to
participants. However, its appearance, timing, and placement posed challenges. Some
participants reported they failed to notice it, especially during their first trial. Furthermore,
its sudden appearance made it hard for participants to rely on. Additionally, a recurrent
concern among participants was the need to switch gaze between the AV and the traffic
light. Finally, participants suggested making the distinction between red and green
clearer.
More unclear, because the pole only appears relatively late. So you're really only focused
on the car that's approaching. And at first, I either saw the pole or I didn't. You see the
12
car approaching sooner, so then you start considering whether you're going to cross or
not. And only then do you see the pole.
Pedestrian lights HUD. The Pedestrian lights HUD also provided a familiar interface. Its
upper-right positioning in their field of view was reported to be both an advantage and a
disadvantage. Specifically, some participants reported that the HUD required them to roll
their eyes, which felt unnatural to them, while others appreciated the constant visibility
and fixed position within their view. Its sudden appearance sometimes led to initial startles
or distractions, and some participants commented that it was obstructive and blocked a
portion of their view. Finally, as with the Fixed pedestrian lights, according to some
participants, the distinction between the red and green signals (lit-up vs. non-lit-up state)
could have been clearer.
“Very clear, indeed. I'm not used to it, so that's why the first one might have been a bit
unclear. Because I was thinking, should I now look or should I clearly know what the car
is going to do? But after that, it's very clear, because it's just close by and you actually
see it right away.”
Planes on vehicle. Participants initially found the Planes on vehicle novel and stated it
took some time to become accustomed to it. The green signal and icon were perceived
as clear. However, there was some uncertainty regarding the red signal, with participants
unsure whether the AV would stop. Many felt that the Planes on vehicle improved trust
because the communication came directly from the AV. One colour blind person,
however, did not immediately understand the meaning of this interface. Furthermore,
some participants remarked that in hypothetical situations involving multiple vehicles, the
effectiveness of the interface might decrease. It was suggested that, for optimal
effectiveness in busier scenarios, all cars might need to adopt such an interface.
I don't know, it made me a bit nervous, because I thought okay... It seems very illogical
that there's a very large sign in front of the car.
“Better than the previous one. I liked that you only have to pay attention to one thing. So
you just have to look at the car and it's clear whether you can cross or not. Instead of
having to look somewhere else for a signal. It was just clearly on the car.”
Finally, a recurring topic regarding the timing of the AR interfaces was that they were activated
somewhat late (even though this was still 0.8 s before the vehicle began to brake). Participants
suggested that the AR interfaces could appear earlier, possibly in a default state, so that they
would know where to look in advance to make their crossing decision. Another theme that
emerged from the interviews is that participants tended to verify the advice of the AR interfaces
with the movement of the vehicle.
Discussion
This study tested four AR interfaces for supporting the interaction with AVs among 28 human
participants, complemented with a baseline condition without AR interface. For the research, we
used actual AR in an outdoor environment, in contrast to previous research that used AR-in-VR.
In our paradigm, participants stood outside and had to press a button as long as they felt safe to
cross. The experiment was set up so that the AV was virtual; this way, we could offer the same
vehicle movement and AR activation timings in every trial.
Some of the results did not match our previous studies. Specifically, a previous experiment in a
CAVE environment (Tabone et al., 2023b) and a large-scale online survey (Tabone et al., 2023a)
showed that among the AR interfaces tested, the world-locked Fixed pedestrian lights and head-
locked Pedestrian lights HUD yielded relatively high intuitiveness ratings while the vehicle-locked
Planes on vehicle received lower ratings, though still net positive. However, in the current study,
the intuitiveness of the Planes on vehicles interface was rated highly, with average scores of 5.5
(SD = 1.6) for non-yielding AVs and 5.9 (SD = 1.2) for yielding AVs on a scale of 1 to 7, whereas
the Fixed pedestrian lights interface received lower ratings, at 4.7 (SD = 1.7) for non-yielding AVs
and 5.5 (SD = 1.2) for yielding AVs, respectively (see Figures 8d and 8e).
13
Although the relative ratings of the AR interfaces did not immediately correspond with prior
research, the information-processing mechanisms did. For example, previous online research
using animated video clips (Tabone et al., 2023a) and research using a virtual pedestrian
simulator (Tabone & De Winter, 2023) also found that a zebra crossing combined with the colour
red can cause confusion, and that walls of the Virtual fence may obstruct the view of other road
users. Additionally, the fact that there can be an egocentric vs. exocentric perspective confusion
when a car emits a red signal is also known (Bazilinskyy et al., 2020), and the problem that world-
locked interfaces, like the Fixed pedestrian lights in our case, cause divided attention, has also
already been documented (e.g., Peereboom et al., 2023). The explanation for the relatively poor
performance of the Fixed pedestrian lights in the present study seems to lie in more practical
factors: We had placed it relatively far away from the participant, at 14 m. Furthermore, for most
of the participants, an orange-coloured aerial work platform was present, located behind the Fixed
pedestrian lights. This visual clutter could have further increased the difficulty in identifying the
traffic light and discerning its status. This research thus shows that basic design decisions related
to salience can have large effects on the effectiveness of AR interfaces.
The Pedestrian lights HUD also suffered from ‘practical limitations’; it was positioned slightly off-
centre, leading to some discomfort as individuals had to adjust their gaze (see Plabst et al., 2022,
for a similar phenomenon). Note that participants could only glance at the HUD if they rotated
their eyes, as the HUD always followed the user’s head movement. The icons of the traffic light
were not particularly salient in their illuminated ‘on’ state compared to their non-illuminated ‘off’
state, causing some participants to glance at the traffic light HUD instead of relying purely on their
peripheral vision. Positioning the HUD towards a more central position would likely increase the
risk of occlusion of relevant objects, which are typically centrally located in users’ fields of view.
Participants experienced fairly low sickness scores, which differs from the higher scores observed
in Peereboom et al.’s (2023) HUD study for pedestrians in VR. This difference may be attributed
to our HUD being placed at a larger virtual distance (1.0 m, compared to their 0.36 m in
Peereboom et al.), reducing the accommodation-vergence conflict. However, it is anticipated that
poorer task performance and/or nausea can occur in more dynamic situations, where the AR
information is not locked to, or embedded in, the world. This concerns the use of HUDs while the
user is walking through the environment, resulting in a moving background relative to static AR
stimuli in the field of vision (Fukushima et al., 2020) or when standing still while the AR stimuli
move in the field of view (Kaufeld et al., 2022). For future research, we suggest creating a HUD
that does not require the user’s direct focus. For instance, the onset of a more bright coloured
light, which can be unambiguously identified through peripheral vision, might offer a more
comfortable experience for the user (e.g., Chaturvedi et al., 2019).
The Fixed pedestrian lights had the disadvantage that it required divided attention. This same
disadvantage was mentioned in earlier experimental research in a CAVE environment, where
participants took a relatively long time to visually identify the traffic lights (Tabone et al., 2023b).
It might be the case that in the current real-world study, the task of identifying the traffic lights and
the AV was even more challenging than in VR, because there is more visual clutter in the real
world, such as other pedestrians and parked vehicles.
Indeed, we noticed that participants seemed to have difficulty with the experiment. For example,
some participants had difficulty localising the attention-attractor circles or had to be reminded to
look into these circles; participants also occasionally had to be reminded to press the response
button if they felt safe to cross. One explanation for this forgetting is that we did not explicitly
provide ‘press now’ instructions before each trial (e.g., Peereboom et al., 2023). Another
explanation for the low press percentages is that the button, which operated through infrared light,
might not have worked reliably in the outdoor environment. It was observed that when participants
held the button in front of their body, instead of next to their body, missing values could result1.
The button press percentages, as shown in Figure 7, resemble online crowdsourcing research on
pedestrian-crossing decisions (e.g., Bazilinskyy et al., 2020, 2021; Sripada et al., 2021), where a
portion of crowdworkers were apparently inattentive. Another possible explanation for the poor
quality of button press data is that mental demands are higher with AR in the real world compared
to AR-in-VR, because the real world is more cluttered, and the participant has more to keep track
1 Excluding trials where no button press was recorded is not advisable. This could introduce bias, as it is unclear whether
participants genuinely did not feel safe crossing or if there was a technical issue.
14
of, including perceiving real objects as well as maintaining postural stability and safety in the
physical world. In turn, these observations suggest that simplicity and clarity of AR communication
are likely even more important than in virtual reality or online experiments.
A limitation of our study is that it used simulated AVs; in future traffic, actual AVs would need to
communicate their stopping intentions wirelessly to the pedestrian’s AR headset. Furthermore,
instead of outside-in tracking using base stations that emit infrared light, there would likely need
to be a form of inside-out tracking, where the pedestrian’s headset detects objects in the
environment (e.g., Bhakar et al., 2022). Although the Varjo XR-3 is a state-of-the-art AR device,
there were some technical hiccups, such as a small number of occasions of loss of tracking, which
caused VR objects to display an abrupt rotation, leading to some confusion or disorientation.
Moreover, while the headset offered a large FOV, it was still more limited than natural vision
without a headset. Finally, in the current study, only a single AV came from the right. Using
multiple vehicles would increase realism, something that is especially relevant for vehicle-locked
AR interfaces. On the other hand, the environment was realistic, with university employees and
students walking around occasionally, as well as some maintenance vehicles being present.
Adding false positives, like a green-coloured signal in combination with an AV that does not stop,
could be useful to investigate (over-)reliance on the AR interfaces (see also Holländer et al., 2019;
Kaleefathullah et al., 2022).
Conclusion
In this study, different AR interfaces for AV-pedestrian interactions were assessed in an outdoor
setting, distinguishing it from previous online and AR-in-VR research. The results showed that
having an AR interface was generally preferred to no AR interface. Additionally, the results of
post-trial interviews replicated previous information-processing-related findings, such as that a
red-coloured surface can be confusing, since this cue can pertain to the AV or to the pedestrian.
The experiment also showed that visual attention mechanisms are of key importance, with a
world-locked traffic light causing challenges since pedestrians need to distribute attention
between the approaching AV and the traffic light. Our findings also demonstrated the importance
of practical design considerations, such as placement and salience, in determining the
effectiveness of AR interfaces. Our results and observations further suggest that participants
found the present task challenging, which, given the complex and cluttered nature of the real
world as opposed to virtual environments, points to the need for simplicity and clarity in AR-based
communication. Future research is thus advised to implement simple AR solutions, such as
through a single coloured light visible in the periphery. Finally, it is recommended to explore forms
of AR for pedestrians where the pedestrian can move around wirelessly. Such research could be
a step towards AR for vulnerable road users, with real application.
Acknowledgements
This project has received funding from the European Union’s Horizon 2020 Research and
Innovation Programme under the Marie Skłodowska-Curie Grant Agreement No. 860410 (project
name: SHAPE-IT).
References
Bazilinskyy, P., Dodou, D., & De Winter, J. C. F. (2020). External Human-Machine Interfaces:
Which of 729 colors is best for signaling ‘Please (do not) cross’? Proceedings of the IEEE
International Conference on Systems, Man and Cybernetics, Toronto, Canada, 3721–3728.
https://doi.org/10.1109/SMC42975.2020.9282998
Bazilinskyy, P., Kooijman, L., Dodou, D., & De Winter, J. C. F. (2021). How should external
Human-Machine Interfaces behave? Examining the effects of colour, position, message,
activation distance, vehicle yielding, and visual distraction among 1,434 participants. Applied
Ergonomics, 95, 103450. https://doi.org/10.1016/j.apergo.2021.103450
Berge, S. H., Hagenzieker, M., Farah, H., & De Winter, J. (2022). Do cyclists need HMIs in
future automated traffic? An interview study. Transportation Research Part F: Traffic
Psychology and Behaviour, 84, 33–52. https://doi.org/10.1016/j.trf.2021.11.013
Bhakar, S., Bhatt, D. P., Dhaka, V. S., & Sarma, Y. K. (2022). A review on classifications of
tracking systems in augmented reality. Journal of Physics: Conference Series (Vol. 2161,
No. 1, p. 012077). IOP Publishing. http://doi.org/10.1088/1742-6596/2161/1/012077
15
Bindschädel, J., Krems, I., & Kiesel, A. (2022). Two-step communication for the interaction
between automated vehicles and pedestrians. Transportation Research Part F: Traffic
Psychology and Behaviour, 90, 136–150. https://doi.org/10.1016/j.trf.2022.08.016
Bokc, T., Maurer, M., & Farber, G. (2007). Validation of the vehicle in the loop (vil); a milestone
for the simulation of driver assistance systems. Proceedings of the 2007 IEEE Intelligent
Vehicles Symposium, Istanbul, Turkey, 612–617. https://doi.org/10.1109/IVS.2007.4290183
Bos, J. E. (2015). Less sickness with more motion and/or mental distraction. Journal of
Vestibular Research: Equilibrium and Orientation, 25, 23–33. https://doi.org/10.3233/ves-
150541
Butenuth, M., Kallweit, R., & Prescher, P. (2017). Vehicle-in-the-loop real-world vehicle tests
combined with virtual scenarios. ATZ Worldwide, 119, 52–55.
https://doi.org/10.1007/s38311-017-0082-4
Chaturvedi, I., Bijarbooneh, F. H., Braud, T., & Hui, P. (2019). Peripheral vision: a new killer app
for smart glasses. Proceedings of the 24th International Conference on Intelligent User
Interfaces, Marina del Ray, CA, 625–636. https://doi.org/10.1145/3301275.3302263
Colley, M., Walch, M., & Rukzio, E. (2020). Unveiling the lack of scalability in research on
external communication of autonomous vehicles. Extended abstracts of the 2020 CHI
Conference on Human Factors in Computing Systems, Honolulu, HI.
https://doi.org/10.1145/3334480.3382865
De Clercq, K., Dietrich, A., Núñez Velasco, P., De Winter, J. C. F., & Happee, R. (2019).
External human-machine interfaces on automated vehicles: Effects on pedestrian crossing
decisions. Human Factors, 61, 1353–1370. https://doi.org/10.1177/0018720819836343
Dey, D., Ackermans, S., Martens, M., Pfleging, B., & Terken, J. (2022). Interactions of
automated vehicles with road users. In A. Riener, M. Jeon, & I. Alvarez (Eds.), User
experience design in the era of automated driving (pp. 533–581). Cham: Springer.
https://doi.org/10.1007/978-3-030-77726-5_20
Dey, D., Holländer, K., Berger, M., Eggen, B., Martens, M., Pfleging, B., & Terken, J. (2020).
Distance-dependent eHMIs for the interaction between automated vehicles and pedestrians.
Proceedings of the 12th International Conference on Automotive User Interfaces and
Interactive Vehicular Applications, Virtual Event DC, 192–204.
https://doi.org/10.1145/3409120.3410642
Epke, M. R., Kooijman, L., & De Winter, J. C. F. (2021). I see your gesture: a vr-based study of
bidirectional communication between pedestrians and automated vehicles. Journal of
Advanced Transportation, 2021, 5573560. https://doi.org/10.1155/2021/5573560
Feng, Y., Yu, C., Xu, S., Liu, H. X., & Peng, H. (2018). An augmented reality environment for
connected and automated vehicle testing and evaluation. Proceedings of the 2018 IEEE
Intelligent Vehicles Symposium, Changshu, China, 1549–1554.
https://doi.org/10.1109/IVS.2018.8500545
Final Form Studio. (2021). Sedan car - 01 [3D land vehicle model]. Unity Asset Store.
https://assetstore.unity.com/packages/3d/vehicles/land/sedan-car-01-190629
Franke, T., Attig, C., & Wessel, D. (2019). A personal resource for technology interaction:
development and validation of the affinity for technology interaction (ATI) scale. International
Journal of Human–Computer Interaction, 35, 456–467.
https://doi.org/10.1080/10447318.2018.1456150
Fukushima, S., Hamada, T., & Hautasaari, A. (2020). Comparing world and screen coordinate
systems in optical see-through head-mounted displays for text readability while walking.
Proceedings of the 2020 IEEE International Symposium on Mixed and Augmented Reality,
649–658, Porto de Galinhas, Brazil. https://doi.org/10.1109/ISMAR50242.2020.00093
Gelbal, S. Y., Cantas, M. R., Guvenc, B. A., Guvenc, L., Surnilla, G., & Zhang, H. (2023). Mobile
safety application for pedestrians. arXiv. https://doi.org/10.48550/arXiv.2305.17575
Google. (2020). A new sense of direction with Live View.
https://blog.google/products/maps/new-sense-direction-live-view
Google Earth. (2023). Top-down view of part of the TU Delft Campus.
https://earth.google.com/web/@52.00008632,4.37098049,0.35726392a,80.22226634d,35y,3
36.33134285h,0t,0r/data=OgMKATA
Hasan, R., & Hasan, R. (2022). Pedestrian safety using the Internet of Things and sensors:
Issues, challenges, and open problems. Future Generation Computer Systems, 134, 187–
203. https://doi.org/10.1016/j.future.2022.03.036
Hesenius, M., Börsting, I., Meyer, O., & Gruhn, V. (2018). Don’t panic! Guiding pedestrians in
autonomous traffic with augmented reality. Proceedings of the 20th International Conference
16
on Human-Computer Interaction with Mobile Devices and Services Adjunct, Barcelona,
Spain, 261–268. https://doi.org/10.1145/3236112.3236148
Holländer, K., Wintersberger, P., & Butz, A. (2019). Overtrust in external cues of automated
vehicles: An experimental investigation. Proceedings of the 11th International Conference on
Automotive User Interfaces and Interactive Vehicular Applications, Utrecht, The Netherlands,
211–221. https://doi.org/10.1145/3342197.3344528
HTC VIVE. (2019). SteamVR Base Station 2.0. https://www.vive.com/eu/accessory/base-
station2
Hussain, K. F., Radwan, E., & Moussa, G. S. (2013). Augmented reality experiment: drivers'
behavior at an unsignalized intersection. IEEE Transactions on Intelligent Transportation
Systems, 14, 608–617. https://doi.org/10.1109/TITS.2012.2226239
Ishihara, S. (1917). Tests for colour-blindness. Tokyo: Kanehara Shuppan Co., Ltd.
Jungmann, A., Lang, C., Pinsker, F., Kallweit, R., Taubenreuther, M., & Butenuth, M. (2020).
Artificial intelligence for automated driving – quo vadis? In T. Bertram (Ed.), Automatisiertes
Fahren 2019: Von der Fahrerassistenz zum autonomen Fahren 5. Internationale ATZ-
Fachtagung (pp. 117–134). Wiesbaden: Springer. https://doi.org/10.1007/978-3-658-27990-
5_11
Kaleefathullah, A. A., Merat, N., Lee, Y. M., Eisma, Y. B., Madigan, R., Garcia, J., & De Winter,
J. C. F. (2022). External Human-Machine Interfaces can be misleading: An examination of
trust development and misuse in a CAVE-based pedestrian simulation environment. Human
Factors, 64, 1070–1085. https://doi.org/10.1177/0018720820970751
Kamalasanan, V., Mukbil, A., Sester, M., & Müller, J. P. (2022). Mixed reality agent-based
framework for pedestrian-cyclist interaction. Proceedings of the 2022 IEEE International
Symposium on Mixed and Augmented Reality Adjunct, Singapore, 363–368.
https://doi.org/10.1109/ISMAR-Adjunct57072.2022.00079
Kang, Y., Choi, S., An, E., Hwang, S., & Kim, S. (2023). Designing virtual agent human–
machine interfaces depending on the communication and anthropomorphism levels in
augmented reality. Proceedings of the 15th International Conference on Automotive User
Interfaces and Interactive Vehicular Applications, Ingolstadt, Germany, 191–201.
https://doi.org/10.1145/3580585.3606460
Kaufeld, M., Mundt, M., Forst, S., & Hecht, H. (2022). Optical see-through augmented reality
can induce severe motion sickness. Displays, 74, 102283.
https://doi.org/10.1016/j.displa.2022.102283
Lakhdhir, S., Somanath, S., & Sharlin, E. (2023). Wearing awareness: Designing pedestrian-
wearables for interactions with autonomous vehicles. Extended Abstracts of the 2023 CHI
Conference on Human Factors in Computing Systems, Hamburg, Germany.
https://doi.org/10.1145/3544549.3585655
Lebeck, K., Ruth, K., Kohno, T., & Roesner, F. (2017). Securing augmented reality output.
Proceedings of the 2017 IEEE Symposium on Security and Privacy, San Jose, CA, 320–337.
https://doi.org/10.1109/SP.2017.13
Lingam, S. N., De Winter, J., Dong, Y., Tsapi, A., Van Arem, B., & Farah, H. (2023). eHMI on
the vehicle or on the infrastructure? A driving simulator study. ResearchGate.
https://www.researchgate.net/publication/362751474_eHMI_on_the_Vehicle_or_on_the_Infr
astructure_A_Driving_Simulator_Study
Malik, J., Kim, N.-Y., Parr, M. D. N., Kearney, J. K., Plumert, J. M., & Rector, K. (2023). Do
simulated augmented reality overlays influence street-crossing decisions for non-mobility-
impaired older and younger adult pedestrians? Human Factors, 00187208231151280.
https://doi.org/10.1177/00187208231151280
Maruhn, P., Dietrich, A., Prasch, L., & Schneider, S. (2020). Analyzing pedestrian behavior in
augmented reality — Proof of concept. Proceedings of the 2020 IEEE Conference on Virtual
Reality and 3D User Interfaces (VR), Atlanta, GA.
https://doi.org/10.1109/vr46266.2020.00051
Morey, R. D. (2008). Confidence intervals from normalized data: A correction to Cousineau
(2005). Tutorials in Quantitative Methods for Psychology, 4, 61–64.
https://doi.org/10.20982/tqmp.04.2.p061
Paiva, S., Ahad, M. A., Tripathi, G., Feroz, N., & Casalino, G. (2021). Enabling technologies for
urban smart mobility: Recent trends, opportunities and challenges. Sensors, 21, 2143.
https://doi.org/10.3390/s21062143
Peereboom, J., Tabone, W., Dodou, D., & de Winter, J. (2023). Head-locked, world-locked, or
conformal diminished-reality? An examination of different AR solutions for pedestrian safety
17
in occluded scenarios. ResearchGate.
https://www.researchgate.net/publication/371509780_Head-locked_world-
locked_or_conformal_diminished-
reality_An_examination_of_different_AR_solutions_for_pedestrian_safety_in_occluded_sce
narios
Plabst, L., Oberdörfer, S., Ortega, F. R., & Niebling, F. (2022). Push the red button: Comparing
notification placement with augmented and non-augmented tasks in AR. Proceedings of the
2022 ACM Symposium on Spatial User Interaction, Online, CA.
https://doi.org/10.1145/3565970.3567701
Pratticò, F. G., Lamberti, F., Cannavò, A., Morra, L., & Montuschi, P. (2021). Comparing state-
of-the-art and emerging augmented reality interfaces for autonomous vehicle-to-pedestrian
communication. IEEE Transactions on Vehicular Technology, 70, 1157–1168.
https://doi.org/10.1109/TVT.2021.3054312
Qualtrics XM. (2023). Qualtrics. https://www.qualtrics.com
Rauschnabel, P. A., Felix, R., Hinsch, C., Shahab, H., & Alt, F. (2022). What is XR? Towards a
framework for augmented and virtual reality. Computers in Human Behavior, 133, 107289.
https://doi.org/10.1016/j.chb.2022.107289
Rodríguez Palmeiro, A. R., van der Kint, S., Vissers, L., Farah, H., de Winter, J. C., &
Hagenzieker, M. (2018). Interaction between pedestrians and automated vehicles: A Wizard
of Oz experiment. Transportation Research Part F: Traffic Psychology and Behaviour, 58,
1005–1020. https://doi.org/10.1016/j.trf.2018.07.020
Sheridan, T. B. (2016). Recollections on presence beginnings, and some challenges for
augmented and virtual reality. Presence: Teleoperators and Virtual Environments, 25, 75–77.
https://doi.org/10.1162/PRES_e_00247
Sripada, A., Bazilinskyy, P., & De Winter, J. C. F. (2021). Automated vehicles that communicate
implicitly: Examining the use of lateral position within the lane. Ergonomics, 64, 1416–1428.
https://doi.org/10.1080/00140139.2021.1925353
Tabone, W., & De Winter, J. C. F. (2023). Using ChatGPT for human-computer interaction: A
primer. Royal Society Open Science, 10, 231053. https://doi.org/10.1098/rsos.231053
Tabone, W., De Winter, J. C. F., Ackermann, C., Bärgman, J., Baumann, M., Deb, S.,
Emmenegger, C., Habibovic, A., Hagenzieker, M., Hancock, P. A., Happee, R., Krems, J.,
Lee, J. D., Martens, M., Merat, N., Norman, D. A., Sheridan, T. B., & Stanton, N. A. (2021a).
Vulnerable road users and the coming wave of automated vehicles: Expert perspectives.
Transportation Research Interdisciplinary Perspectives, 9, 100293.
https://doi.org/10.1016/j.trip.2020.100293
Tabone, W., Happee, R., Garcia, J., Lee, Y. M., Lupetti, M. C., Merat, N., & De Winter, J. C. F.
(2023a). Augmented reality interfaces for pedestrian-vehicle interactions: An online study.
Transportation Research Part F: Traffic Psychology and Behaviour, 94, 170–189.
https://doi.org/10.1016/j.trf.2023.02.005
Tabone, W., Happee, R., Yang, Y., Sadraei, E., García, J., Lee, Y. M., Merat, N., & De Winter,
J. C. F. (2023b). Immersive insights: Evaluating augmented reality interfaces for pedestrians
in a CAVE-based experiment. ResearchGate.
https://www.researchgate.net/publication/370160064_Immersive_Insights_Evaluating_Augm
ented_Reality_Interfaces_for_Pedestrians_in_a_CAVE-Based_Experiment
Tabone, W., Lee, Y. M., Merat, N., Happee, R., & De Winter, J. C. F. (2021b). Towards future
pedestrian-vehicle interactions: Introducing theoretically-supported AR prototypes. 13th
International ACM Conference on Automotive User Interfaces and Interactive Vehicular
Applications, 209–218, Leeds, UK. https://doi.org/10.1145/3409118.3475149
Tong, Y., Jia, B., & Bao, S. (2021). An augmented warning system for pedestrians: user
interface design and algorithm development. Applied Sciences, 11, 7197.
https://doi.org/10.3390/app11167197
Tran, T. T. M., Parker, C., & Tomitsch, M. (2023). Scoping out the scalability issues of
autonomous vehicle-pedestrian interaction. Proceedings of the 15th International
Conference on Automotive User Interfaces and Interactive Vehicular Applications, Ingolstadt
Germany, 167–177. https://doi.org/10.1145/3580585.3607167
Tran, T. T. M., Parker, C., Wang, Y., & Tomitsch, M. (2022). Designing wearable augmented
reality concepts to support scalability in autonomous vehicle-pedestrian interaction. Frontiers
in Computer Science, 4. https://doi.org/10.3389/fcomp.2022.866516
Transparency Market Research. (2023). Automotive Head-up Display (HUD) market.
https://www.transparencymarketresearch.com/automotive-head-up-display-market.html
18
Troel-Madec, M., Boissieux, L., Borkoswki, S., Vaufreydaz, D., Alaimo, J., Chatagnon, S., &
Spalanzani, A. (2019). eHMI positioning for autonomous vehicle/pedestrians interaction.
Adjunct Proceedings of the 31st Conference on l’Interaction Homme-Machine, Grenoble,
France. https://doi.org/10.1145/3366551.3370340
United Nations. (2020). United Nations A/RES/74/299 (A/RES/74/299). Resolution adopted by
the General Assembly on 31 August 2020. https://documents-dds-
ny.un.org/doc/UNDOC/GEN/N20/226/30/PDF/N2022630.pdf?OpenElement
Unity. (2022). Unity 2021.3.13. https://unity.com/releases/editor/whats-new/2021.3.13
Van der Laan, J. D., Heino, A., & De Waard, D. (1997). A simple procedure for the assessment
of acceptance of advanced transport telematics. Transportation Research Part C: Emerging
Technologies, 5, 1–10. https://doi.org/10.1016/s0968-090x(96)00025-3
Varjo. (2023). Varjo XR-3 - The industry’s highest resolution mixed reality headset.
https://varjo.com/products/xr-3
Varjo Developer. (2023). Getting started with Varjo XR Plugin for Unity.
https://developer.varjo.com/docs/unity-xr-sdk/getting-started-with-varjo-xr-plugin-for-unity
Varjo Technologies. (2023). Mixed reality – Varjo.com. https://varjo.com/use-center/get-to-
know-your-headset/mixed-reality/#reference-marker
Wilbrink, M., Burgdorf, F. K. G., & Oehl, M. (2023). AR designs for eHMI–Communication
between automated vehicles and pedestrians using augmented reality. Proceedings of the
8th Humanist Conference, Berlin, Germany.
World Health Organization. (2018). Global status report on road safety 2018: Summary.
https://apps.who.int/iris/handle/10665/277370
Zhang, X., Wu, X., Cools, R., Simeone, A. L., & Gruenefeld, U. (2023). ARcoustic: A mobile
augmented reality system for seeing out-of-view traffic. Proceedings of the 15th International
Conference on Automotive User Interfaces and Interactive Vehicular Applications, Ingolstadt,
Germany, 178–190. https://doi.org/10.1145/3580585.3606461
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Many collisions between pedestrians and cars are caused by poor visibility, such as occlusion by a parked vehicle. Augmented reality (AR) could help to prevent this problem, but it is unknown to what extent the augmented information needs to be embedded into the world. In this virtual reality experiment with a head-mounted display (HMD), 28 participants were exposed to AR designs, in a scenario where a vehicle approached from behind a parked vehicle. The experimental conditions included a head-locked live video feed of the occluded region, meaning it was fixed in a specific location within the view of the HMD (VideoHead), a world-locked video feed displayed across the street (VideoStreet), and two conformal diminished reality designs: a see-through display on the occluding vehicle (VideoSeeThrough) and a solution where the occluding vehicle has been made semi-transparent (TransparentVehicle). A Baseline condition without augmented information served as a reference. Additionally, the VideoHead and VideoStreet conditions were each tested with and without the addition of a guiding arrow indicating the location of the approaching vehicle. Participants performed 42 trials, 6 per condition, during which they had to hold a key when they felt safe to cross. The keypress percentages and responses from additional questionnaires showed that the diminished-reality TransparentVehicle and VideoSeeThrough designs came out most favourably, while the VideoHead solution caused some discomfort and dissatisfaction. An analysis of head yaw angle showed that VideoHead and VideoStreet caused divided attention between the screen and the approaching vehicle. The use of guiding arrows did not contribute demonstrable added value. AR designs with a high level of local embeddedness are beneficial for addressing occlusion problems when crossing. However, the head-locked solutions should not be immediately dismissed because, according to the literature, such solutions can serve tasks where a salient warning or instruction is beneficial.
Conference Paper
Full-text available
Autonomous vehicles (AVs) may use external interfaces, such as LED light bands, to communicate with pedestrians safely and intuitively. While previous research has demonstrated the effectiveness of these interfaces in simple traffic scenarios involving one pedestrian and one vehicle, their performance in more complex scenarios with multiple road users remains unclear. The scalability of AV external communication has therefore attracted increasing attention, prompting the need for further investigation. This scoping review synthesises information from 54 papers to identify seven key scalability issues in multi-vehicle and multi-pedestrian environments, with Clarity of Recipients, Information Overload, and Multi-Lane Safety emerging as the most pressing concerns. To guide future research in scalable AV-pedestrian interactions, we propose high-level design directions focused on three communication loci: vehicle, infrastructure, and pedestrian. Our work contributes the groundwork and a roadmap for designing simplified, coordinated, and targeted external AV communication, ultimately improving safety and efficiency in complex traffic scenarios.
Article
Full-text available
ChatGPT could serve as a tool for text analysis within the field of Human–Computer Interaction, though its validity requires investigation. This study applied ChatGPT to: (1) textbox questionnaire responses on nine augmented-reality interfaces, (2) interview data from participants who experienced these interfaces in a virtual simulator, and (3) transcribed think-aloud data of participants who viewed a real painting and its replica. Using a hierarchical approach, ChatGPT produced scores or summaries of text batches, which were then aggregated. Results showed that (1) ChatGPT generated sentiment scores of the interfaces that correlated extremely strongly (r > 0.99) with human rating scale outcomes and with a rule-based sentiment analysis method (criterion validity). Additionally, (2) by inputting automatically transcribed interviews to ChatGPT, it provided meaningful meta-summaries of the qualities of the interfaces (face validity). One meta-summary analysed in depth was found to have substantial but imperfect overlap with a content analysis conducted by an independent researcher (criterion validity). Finally, (3) ChatGPT’s summary of the think-aloud data highlighted subtle differences between the real painting and the replica (face validity), a distinction corresponding with a keyword analysis (criterion validity). In conclusion, our research indicates that, with appropriate precautions, ChatGPT can be used as a valid tool for analysing text data.
Preprint
Full-text available
Vulnerable Road User (VRU) safety has been an important issue throughout the years as corresponding fatality numbers in traffic have been increasing each year. With the developments in connected vehicle technology, there are new and easier ways of implementing Vehicle to Everything (V2X) communication which can be utilized to provide safety and early warning benefits for VRUs. Mobile phones are one important point of interest with their sensors being increased in quantity and quality and improved in terms of accuracy. Bluetooth and extended Bluetooth technology in mobile phones has enhanced support to carry larger chunks of information to longer distances. The work we discuss in this paper is related to a mobile application that utilizes the mobile phone sensors and Bluetooth communication to implement Personal Safety Message (PSM) broadcast using the SAE J2735 standard to create a Pedestrian to Vehicle (P2V) based safety warning structure. This implementation allows the drivers to receive a warning on their mobile phones and be more careful about the pedestrian intending to cross the street. As a result, the driver has much more time to safely slow down and stop at the intersection. Most importantly, thanks to the wireless nature of Bluetooth connection and long-range mode in Bluetooth 5.0, most dangerous cases such as reduced visibility or No-Line-of-Sight (NLOS) conditions can be remedied.
Article
Full-text available
Augmented Reality (AR) technology could be utilised to assist pedestrians in navigating safely through traffic. However, whether potential users would understand and use such AR solutions is currently unknown. Nine novel AR interfaces for pedestrian-vehicle communication, previously developed using an experience-based design method, were evaluated through an online questionnaire study completed by 992 respondents in Germany, the Netherlands, Norway, Sweden, and the United Kingdom. The AR indicated whether it was safe to cross the road in front of an approaching automated vehicle. Each interface was rated for its intuitiveness and convincingness, aesthetics, and usefulness. Moreover, comments were collected for qualitative analysis. The results indicated that interfaces that employed traditional design elements from existing traffic, and head-up displays, received the highest ratings overall. Statistical results also showed that there were no significant effects of country, age, and gender on interface acceptance. Thematic analysis of the textual comments offered detail on each interface design's stronger and weaker points, and revealed unintended effects of certain designs. In particular, some of the interfaces were commented on as being dangerous or scary, or were criticised that they could be misinterpreted in that they signal that something is wrong with the vehicle, or that they could occlude the view of the vehicle. The current findings highlight the limitations of experience-based design, and the importance of applying legacy design principles and involving target users in design and evaluation. Future research should be conducted in scenarios in which pedestrians actually interact with approaching vehicles.
Article
Full-text available
Objective: This study used a virtual environment to examine how older and younger pedestrians responded to simulated augmented reality (AR) overlays that indicated the crossability of gaps in a continuous stream of traffic. Background: Older adults represent a vulnerable group of pedestrians. AR has the potential to make the task of street-crossing safer and easier for older adults. Method: We used an immersive virtual environment to conduct a study with age group and condition as between-subjects factors. In the control condition, older and younger participants crossed a continuous stream of traffic without simulated AR overlays. In the AR condition, older and younger participants crossed with simulated AR overlays signaling whether gaps between vehicles were safe or unsafe to cross. Participants were subsequently interviewed about their experience. Results: We found that participants were more selective in their crossing decisions and took safer gaps in the AR condition as compared to the control condition. Older adult participants also reported reduced mental and physical demand in the AR condition compared to the control condition. Conclusion: AR overlays that display the crossability of gaps between vehicles have the potential to make street-crossing safer and easier for older adults. Additional research is needed in more complex real-world scenarios to further examine how AR overlays impact pedestrian behavior. Application: With rapid advances in autonomous vehicle and vehicle-to-pedestrian communication technologies, it is critical to study how pedestrians can be better supported. Our research provides key insights for ways to improve pedestrian safety applications using emerging technologies like AR.
Conference Paper
With the introduction of autonomous vehicles, pedestrians may no longer expect explicit communication from drivers. Despite the anticipated safety benefits of anthropomorphic human–machine interfaces (HMIs) for pedestrian crossings, the impact of different levels of anthropomorphism and communication on pedestrian safety remains insufficiently understood. We proposed a virtualagent (VA) HMI that mimics driver behavior and investigated pedestrians’ preferences through augmented reality (AR) experiments. Eighteen participants made decisions about crossing after receiving cues about the vehicle’s intentions from VA HMIs. Participants preferred the "characterized" VA HMI owing to its aesthetically pleasing design and found the "eye contact + hand gesture" combination to be more easily comprehensible. We found that while the degree of anthropomorphism did not significantly affect pedestrians’ crossing decisions, more explicit communication was helpful. Our study provides empirical evidence regarding users’ experiences of HMI in AR and the effectiveness of VA HMIs that imitate driver communication modes.
Conference Paper
Participating in urban traffic is inherently risky for humans. There-fore, in psychology, behavioural studies have been using Virtual Reality (VR) to simulate and experiment with human behaviour. Safety critical interactions (e.g. conflict, collision or near collision) can be captured from the motion trajectories. However, the motion data in virtual settings is influenced by the modelling software used to create the virtual world, which might fail to capture one-to-one interactions accurately (such as interactions between pedestrians and cyclists in mixed traffic). Our system paper proposes a Pedestrian-in-the-Loop (PIL) Mixed Reality (MR) framework, where mobile virtual cyclist avatars co-exist with humans in a real-world outdoor space. Such a setting can be used to study a pedestrian subject, both viewing and interacting with moving holograms of cyclists in real traffic. The novelty of our approach is modelling virtual avatars as cognitive agents. To achieve this, we integrate agent-based models so that the virtual avatar can sense the environment and interact with the real user participating in the experiments. We demonstrate that this approach could contribute to effectively studying of pedestrian interactions. We also perform an evaluation to quantify the amount of trajectory error for our outdoor framework. For this, we compare the position data of a subject during an experiment to a proven benchmark for indoor motion capture. Additionally, an application of using the framework to demonstrate pedestrian dominance is presented